Unverified50% confidenceFactExact time
Experiments correcting erroneous constraints show that when the model receives an erroneous note prohibiting contact, correcting that note can restore the permitted inquiry behavior, confirming that the model blindly accepts externally assumed constraints.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified系统会试探性地应用补丁，若失败则将错误信息反馈给模型进行修正，构成闭环纠错机制79% similarUnverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉77% similarUnverifiedExperiments show that non-agent methods frequently miss changes requested by users while also disturbing regions that should remain unchanged.74% similarUnverified下游用户对模型的二次微调可能意外破坏原有安全特性73% similarUnverifiedUnder injection bait, weak models treat the injected incorrect numbers as their own calculation results and fabricate verification processes out of thin air, with the explanatory text appearing consistent with genuine verification, making it impossible to detect errors through auditing.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1078898API
curl https://kongchang.com/api/v1/knowledge/claims/1078898MCP
get_claim(id=1078898)